# Yashwardhan Chaudhuri > Researcher → applied scientist **Agents · training · retrieval · evals** Applied ML scientist. Previously Wadhwani AI and IIIT-Delhi. I like to think from **first principles**. I started my journey in research labs where I slept and worked, from no GPUs to external grants for my first workstation. What pulled me toward applied science was the ambition of generating socio-economic value from research, and whether the work created value for someone who needed it. I call it **business aligned R&D**. I can go in as much depth as required if it solves a real problem. ## Contact I take a few calls each month for research collaborations, consulting, and roles where the work ships. - **Email:** yashwardhan@noteweave.io - **Book a call:** https://app.cal.com/yashwardhanc/30min - **LinkedIn:** https://www.linkedin.com/in/yashwardhan-chaudhuri - **Scholar:** https://scholar.google.com/citations?user=VV2eC_gAAAAJ - **Resume:** https://yashwardhan.space/resume/main.pdf ## Work ### Founder · 2025–2026 ## Noteweave > Autonomous R&D lab for research-to-deployment - **Role:** Co-founder (Product + Research) - **Company:** Noteweave - **Period:** 2026 – Present Noteweave was built as an **autonomous R&D lab** for moving research into usable work. The first product shipped in **April 2026** as a **VS Code extension**: search over **3M+ arXiv papers**, full-paper critique with **E3**, and cited planning inside the developer workflow. The stack uses **FastAPI**, **LangGraph**, **Elasticsearch**, **Qdrant**, **MiniLM**, **Gemini grounded search**, and LLM reranking. E3 was evaluated on **100 ICLR 2026 submissions** and recovered **90.2%** of issue-level review points ([arXiv:2605.27072](https://arxiv.org/abs/2605.27072)). Bootstrapped on **Brahma**, a second-hand laptop server, and backed by **Founders Inc. Canopy**. **Focus:** Autonomous R&D, Research agents, Hybrid search, Paper critique, VS Code, Independent lab ## Supported by - Amazon Web Services (Activate) - OpenAI (for startups) - Anthropic (for startups) - Microsoft (for startups) - Founders Inc (Canopy 2026) Plus **~$50k** startup credits, a **$10k OpenAI grant**, and support from **Cursor**, **PostHog**, and **Sarvam**, after bootstrapping on **Brahma**, our second-hand laptop home server. [LeadAgent fans out parallel search and E3 workers into the VS Code sidebar.](https://yashwardhan.space/assets/videos/noteweave_clips/exploration_agents_clip.mp4) ## Overall structure The product covers two R&D steps today: **screen** and **understand**. A user gives a question or requirement; Noteweave searches the literature, critiques the strongest papers, and writes a cited plan in VS Code. The lab extends the same stack toward the third step: **deploy**. **Take home:** **VS Code extension** on a **FastAPI** backend, with search, critique, and project memory tied to the workspace. - **150+** extension + web users - **3M+** arXiv hybrid index - **100** ICLR 2026 eval papers - **90.2%** E3 recall (partial) ![Noteweave platform map](https://yashwardhan.space/assets/images/work/noteweave/platform-architecture.svg?v=2) *Platform map* *VS Code extension, shared literature index, E3 critique, and project memory.* ## Bootstrapped on Brahma Before any cloud credits arrived, the entire product ran on **Brahma**, a second-hand gaming laptop taped to the wall. It held the search index, the models, and the live agent sessions. We indexed arXiv, served real users, and shipped the VS Code extension from that one box in under a month. **Take home:** Running it ourselves first cut cloud spend by ~90% in month one and proved the architecture before we ever paid for scale. - **~90%** cloud savings - **1** laptop server - **<1 mo** to VS Code ship ![Brahma, second-hand ASUS ROG laptop used as Noteweave home server](https://yashwardhan.space/assets/images/work/noteweave/brahma.jpeg) *Brahma, home server built from a second-hand gaming laptop* *Wall-mounted ASUS ROG running the full agent stack before cloud credits. ~90% cloud savings in month one.* ## Search and critique stack Built the search-and-critique path for Noteweave: turn a requirement into typed searches, retrieve papers, rerank them, run **E3** on full text, and write a cited plan. Search finds candidate work, reranking selects what matters, E3 judges reliability, and the writer only assembles after those signals exist. ### Query planner - The planner expands a requirement into typed searches: **method**, **metric**, **failure mode**, **alternative architecture**, **dataset**, **efficiency**, **survey**, and the user's exact wording. - This keeps the search close to the requirement while still covering adjacent papers a researcher would check manually. ### Retrieval - **3M+ arXiv papers** are indexed two ways: **BM25** in Elasticsearch for exact terms and **MiniLM vectors** in Qdrant for semantic matches. **Reciprocal Rank Fusion** merges both lists. - **Gemini grounded search** adds fresh web-grounded recall. **Semantic Scholar** backfills venue and citation metadata, so venue-specific search works from the local cache. ![Noteweave hybrid retrieval and ranking pipeline](https://yashwardhan.space/assets/images/work/noteweave/retrieval-pipeline.svg) *Retrieval and ranking pipeline* *Typed queries, BM25, MiniLM vectors, Gemini grounded search, RRF, LLM rerank, and cited output.* ### Rerank and read - An **LLM reranker** scores title, abstract, venue, and citations against the requirement. Top papers move to full-text reading. - The final plan links back to retrieved papers and E3 critiques, so the reasoning path is visible. ### Retrieval eval - Internal **nDCG@10** eval: grounded search scored ~**0.41** when it returned; local **BM25 + dense** scored ~**0.33** with stable coverage; twelve-provider fan-out scored ~**0.13**. - The shipped default uses grounded search plus the local hybrid fallback. [Coding agents in VS Code](https://yashwardhan.space/assets/videos/noteweave_clips/exploration_agents_clip.mp4) *Search, E3 analysis, and cited writing streamed into VS Code.* ## E3: automated paper critique **E3** is the paper-quality judge. Pass 1 reads the abstract and builds a reviewer lens: **subfield**, **core mechanism**, **critical assumptions**, **expert checkpoints**, and **known failure modes**. Pass 2 reads the full paper and writes a structured review with cited evidence. ### What E3 judges - **Technical summary**, **core claim**, **main risks**, **domain-specific concerns**, **strengths**, **weaknesses**, **forensic deep-dive**, **missing evaluations**, and **sharpest flaw**. - **Dataset and deployment audit**, **productionizability scorecard**, **acceptance recommendation**, and atomic review pointers. Every criticism cites a section, table, equation, or sentence. ### Evaluation - Benchmarked against human peer reviews on **100 ICLR 2026 submissions** (accept and reject outcomes), all post-dating model training cutoffs to rule out memorization, with a blinded judge scoring which substantive issues each source recovered. - E3 recovered **90.2%** of issues, the highest among human, E3, and matched **GPT** and **Claude** baselines. The full protocol and results are in the paper ([arXiv:2605.27072](https://arxiv.org/abs/2605.27072)). ![E3 issue-level backtesting severity chart](https://yashwardhan.space/assets/images/work/noteweave/e3-critique.jpg) *Issue-level backtesting results* *Recall and severity comparison across human, E3, GPT, and Claude review streams.* ![E3 issue-level backtesting protocol](https://yashwardhan.space/assets/images/work/noteweave/e3-protocol.svg?v=2) *E3 evaluation protocol* *Post-cutoff corpus, blinded meta-judge, and 4,598 judged issue rows across four review streams.* ## Traction & funding April 2026: VS Code extension and E3 paper shipped the same month. **Founders Inc. Canopy 2026** residency in San Francisco. **150+ accounts** across the extension and web app; early teams include **Dognosis** and **Audria**. YC Product Hunt application ranked **#17 internationally** and **#22 overall**. ### Programs & credits - **Founder in Residence** at Founders Inc. Canopy; shipped **hybrid search**, **E3**, and the **VS Code extension**. - Startup credits from **AWS Activate**, **Microsoft for Startups**, **Anthropic**, **OpenAI**, and **Grok (xAI)**; **$10k OpenAI Startup Grant**. - Accepted to **YC Startup School**. ## Closing the loop: the autonomous lab The lab extends Noteweave from paper review to experiment runs: propose a hypothesis, run it, critique the result, and choose the next step. **Take home:** **Search + E3** moves from reviewing papers to evaluating experiments. ### Closing the loop - Hypothesis, experiment, critique, next iteration, chained so the agent can carry a line of research forward on its own. - Built on the same search and critique agents that already shipped in the editor. ![Recursive experimentation harness, independent lab preview](https://yashwardhan.space/assets/images/work/noteweave/auto-scientist.jpeg) *Recursive experimentation harness* *Independent lab preview, agents that propose, run, and critique experiments in a closed loop.* ## Impact **VS Code extension** + **FastAPI** service. Hybrid search over **3M+ arXiv papers**, **E3** critique, and project memory. SOC-II Type 2 in progress. - Shipped in VS Code: literature search, paper critique, and cited planning. - Hybrid search over 3M+ arXiv papers combining keyword and meaning-based matching, kept current nightly and filterable by venue. - E3 critique engine: 90.2% recall of real issues against ICLR 2026 peer reviews, highest of all sources ([arXiv:2605.27072](https://arxiv.org/abs/2605.27072)). Backed by Founders Inc. Canopy. ## Sources - [noteweave.io](https://www.noteweave.io) - [VS Code extension](https://marketplace.visualstudio.com/items?itemName=noteweave.noteweave) - [E3 paper (arXiv)](https://arxiv.org/abs/2605.27072) HTML case study: https://yashwardhan.space/work.html?p=noteweave --- ## ResXiv > Search, cited reading, critique, and writing on one shared library - **Role:** Co-founder (Product + Research) - **Company:** ResXiv - **Period:** 2025 – Mar 2026 - **Users from:** Cornell, Princeton, Harvard, Adobe, Google, Aftershoot, IIIT Delhi, CMU, UMD, IIT, and more ResXiv was a browser research workspace: search papers, save them to a project, chat with citations, review drafts, and write in LaTeX from the same library. The backend used **FastAPI**, **PostgreSQL**, **Qdrant**, and **LangGraph**. Search covered **11 scholarly providers** and **280M+ catalog records** through live provider fan-out. Saved PDFs were chunked into Qdrant for cited RAG; drafts and LaTeX documents lived in the project database. The beta reached **300+ MAU** with users from Cornell, Princeton, Harvard, Adobe, Google, and others. ResXiv closed in **March 2026**; the search, critique, and IDE lessons carried into Noteweave. **Focus:** Research search, RAG, LangGraph, Fusion search, LaTeX, FastAPI, PostgreSQL, Qdrant [ResXiv search, eleven scholarly providers in one query.](https://yashwardhan.space/assets/videos/resxiv_clips/fusion_search_clip.mp4) ## Overall structure ResXiv kept research work in one project library. **Search** found papers, **RAG chat** answered from saved PDFs, **draft critique** checked citations and risks, and the **LaTeX editor** wrote against the same project records. **Take home:** **FastAPI** backend with **PostgreSQL** for projects and documents, **Qdrant** for PDF chunks, and **LangGraph** for research-agent routing. - **7** AI products - **3** core surfaces - **11** search providers - **300+** peak MAU ![ResXiv platform map](https://yashwardhan.space/assets/images/work/resxiv/system-architecture.svg?v=2) *Platform map* *Search, cited chat, draft critique, and LaTeX writing on one FastAPI backend.* ## Search, RAG, and writing stack The stack had three surfaces: live paper search, cited answers over saved PDFs, and draft / LaTeX writing. Each used the same project library. ### Search - One query fan-outs to **11 providers**: OpenAlex, arXiv, PubMed, bioRxiv, medRxiv, EuropePMC, ClinicalTrials, ISRCTN, PubChem, ChemRxiv, and Hugging Face Papers. - Results are normalized, deduped, filtered by year/provider when needed, and saved into the **PostgreSQL** project library. - Smart search used LLM routing for source choice and query intent. ### Cited chat - Saved PDFs ingest into **Qdrant** as chunks. Paper chat answers from those chunks with mandatory citations. - **LangGraph** routes research questions to search, project papers, paper analysis, web search, or author / field nodes. - X-Ray analysis highlights weak claims and unsupported lines inside a paper. ### Draft and LaTeX - **Draft critique** checks missing citations and risk before submission. - The **LaTeX editor** stores documents as project records, converts between editor schema and TeX, and keeps compile state tied to the same backend. - Notes, papers, drafts, and chats stayed inside one project workspace. ![ResXiv search, RAG, and writing architecture](https://yashwardhan.space/assets/images/work/resxiv/ai-layer.svg?v=1) *Search and RAG stack* *Live scholarly search, PostgreSQL project library, Qdrant PDF index, and LaTeX workspace.* ## Problem The core issue was citation trust. ResXiv answered from saved papers and project documents, so paper chat, draft critique, and research reports could point back to source material. [Search](https://yashwardhan.space/assets/videos/resxiv_clips/fusion_search_clip.mp4) *One query across eleven scholarly indexes with discipline tagging and deduplication.* ## What shipped The beta shipped five user-facing surfaces. ### Search - One query across **280M+** catalog records through live provider fan-out. [Search](https://yashwardhan.space/assets/videos/resxiv_clips/fusion_search_clip.mp4) *Eleven scholarly providers, normalized and deduped into one result list.* ### Research agent - Research question to cited field report using **LangGraph** routing. [Research agent](https://yashwardhan.space/assets/videos/resxiv_clips/research_agent_clip.mp4) *LangGraph routes questions to search, project papers, paper analysis, and web search.* ### Paper chat - Cited answers from saved PDFs in the project library. [Paper chat](https://yashwardhan.space/assets/videos/resxiv_clips/paper_chat_clip.mp4) *Qdrant RAG over ingested paper chunks with mandatory citations.* ### Write workspace - Notes, AI edits, and LaTeX compile in one workspace. [Write workspace](https://yashwardhan.space/assets/videos/resxiv_clips/write_workspace_clip.mp4) *Editor schema, TeX conversion, and compile tied to project records.* ### Draft analysis - Missing citations and draft risks before submission. [Draft analysis](https://yashwardhan.space/assets/videos/resxiv_clips/draft_analysis_clip.mp4) *Draft critique with citation checks and risk summary.* ## Impact Shipped as a production beta. Usage centered on search, cited paper chat, and draft critique. - **300+** monthly active users - **280M+** search catalog - **11** search providers ### Users from - Researchers and engineers from **Cornell**, **Princeton**, **Harvard**, **CMU**, **UMD**, **IIIT Delhi**, **IIT**, and other universities. - Industry users from **Adobe**, **Google**, **Aftershoot**, and more. - Top usage discipline tags: **AI/ML**, **medicine**, and **biology**. ### Traction - **300+ MAU** at peak on the production beta. - Co-founded with **Paridhi Mundra**, who led product and design. Search gateway, agents, RAG, ingestion, and editor were built on the research and engineering side. - Presented at the **Bharat Bodhan AI Conclave**. ![ResXiv booth at Bharat Bodhan AI Conclave](https://yashwardhan.space/assets/images/work/resxiv/resxiv-booth-crop.jpeg) *Bharat Bodhan AI Conclave* *Presenting ResXiv at the conclave after reaching 300+ MAU on the production beta.* ## Team - [Yashwardhan Chaudhuri](https://www.linkedin.com/in/yashwardhan-chaudhuri) (Co-founder (Product + Research)) - [Paridhi Mundra](https://www.linkedin.com/in/paridhi-mundra-1ba31622b) (Co-founder (Product + Design)) ## Impact Shipped as a production beta, then closed March 2026. - Seven products on one project library: search, research agent, RAG chat, draft critique, write workspace, AI-TeX, and ingestion. - 300+ MAU, with users from Cornell, Princeton, Harvard, Adobe, Google, Aftershoot, IIIT Delhi, CMU, UMD, and IIT. - Built on FastAPI, PostgreSQL, Qdrant, LangGraph, live scholarly APIs, and a LaTeX document service. ## Sources - [Full demo (YouTube)](https://www.youtube.com/watch?v=zhQY4xie4BA) HTML case study: https://yashwardhan.space/work.html?p=resxiv --- ### Consulting · 2025–2026 ## DraftNCraft > Patent & technical document discovery - **Role:** Independent Consultant - **Company:** DraftNCraft · Private engagement - **Period:** Sep 2025 – Apr 2026 Patent analysts need both exact claim matches and conceptually related prior art. I designed the search architecture for **1M+ patent records** at Global IP Services and built the search index, dual-index retrieval backbones, and ingestion pipeline for prior-art discovery at scale. **Focus:** Hybrid search, Patent discovery, RapidOCR, Hunyuan-MT, FastAPI, PostgreSQL, vLLM [Hybrid prior-art search over a dual-index patent corpus.](https://yashwardhan.space/assets/videos/hybrid_search_real.mp4) ## Overall structure Built the **backend patent search infrastructure**: canonical records, dual retrieval engines, and indexing for multi-jurisdiction prior art. **Take home:** Shipped the search gateway, ingestion pipeline, and dual-index backbones the product would sit on. - **1M+** records (design target) - **2** retrieval engines - **L40** RapidOCR + MT co-hosted - **Dual index** ES + Qdrant ![DraftNCraft platform map](https://yashwardhan.space/assets/images/work/draftncraft/platform-architecture.svg?v=6) *Platform map* *Find and Index on a shared PostgreSQL, Elasticsearch, and Qdrant backend for multi-jurisdiction prior-art search.* ## Hybrid search Hybrid search runs **lexical and semantic recall in parallel**, merges the result sets, and reranks so analysts see both exact claim matches and conceptually related prior art. Deeper pagination favors the faster lexical path for latency. ### Why two indices - **Lexical search** catches exact claim language, patent numbers, and OCR text from scanned filings. - **Semantic search** catches paraphrases and conceptually related prior art that keyword lookup misses. - Single-index search forces a tradeoff; dual recall keeps both precision and recall. ### Merge and rerank - Both retrieval paths run in parallel, then results merge into one ranked list. - Hits found by **both** paths rank highest; single-path hits stay in the mix at lower weight. - Full filing text and metadata come from the **canonical store**, not the index layers. ### Production search - Hybrid search is the default analyst path; keyword-only and semantic-only modes remain available. - Filters for jurisdiction and patent class sit on top of the merged result set. ![DraftNCraft hybrid search](https://yashwardhan.space/assets/images/work/draftncraft/hybrid-retrieval.svg?v=6) *Hybrid prior-art search* *Parallel lexical and semantic recall, merged and reranked for prior-art discovery.* ## Indexing pipeline Each filing moves through ingest → text extraction → language normalization → canonical write → dual-index build. Query-time hybrid search is a separate path over the same corpus. ### OCR for scanned filings - **RapidOCR** runs on scanned pages where native text extraction is insufficient. - Native PDF text extraction runs first to avoid unnecessary OCR cost and latency. - Benchmarked against GLM OCR on patent layouts; RapidOCR won on accuracy and latency for single-GPU serving. ### Translation - Non-English filings normalize to English before chunking and embedding. - Language detection gates translation so English filings skip the extra step. ### GPU serving - RapidOCR and Hunyuan-MT co-host on one **L40 GPU** with warm-loaded weights at startup. - Heavy indexing jobs serialize on the GPU so OCR and translation do not contend at peak load. ![DraftNCraft indexing pipeline](https://yashwardhan.space/assets/images/work/draftncraft/ingestion-pipeline.svg?v=6) *Indexing pipeline* *Five-step flow from filing PDF to searchable prior-art record.* ## What I built Hands-on work across the search gateway, indexing pipeline, database setup, and OCR/translation serving, designed for **1M+ patent records** across USPTO, EPO, and WIPO corpora. ### Indexing pipeline - End-to-end workflow from filing ingest through canonical write and dual-index build. - Single and bulk indexing paths with status tracking in the canonical store. ### Hybrid search - Parallel lexical + semantic retrieval with merge, rerank, and pagination. - Hybrid, keyword-only, and semantic-only modes over the same indexed corpus. ### Infrastructure - PostgreSQL schema, Elasticsearch mappings, and Qdrant collections for million-scale prior art. - Multi-jurisdiction ingest paths for USPTO, EPO, WIPO, and other patent offices. ## Scale Architecture and core stack for **1M+ patent records** across multi-jurisdiction corpora: hybrid search, indexing pipeline, canonical store, and **RapidOCR** + Hunyuan-MT on one L40. Delivered the **search index** and retrieval backbones that anchor the platform. - **1M+** records (design target) - **Multi-office** USPTO · EPO · WIPO - **Dual index** Elasticsearch + Qdrant - **Delivered** search + indexing API ## Impact Search gateway, indexing pipeline, and dual-index backbones (PostgreSQL + Elasticsearch + Qdrant) designed for 1M+ multi-jurisdiction patent records. - Hybrid search: parallel lexical + semantic recall, merged and reranked for prior-art discovery - Indexing pipeline: ingest → extract/OCR → translate → canonical store → dual-index build - RapidOCR and Hunyuan-MT on one L40 for million-scale filing ingestion HTML case study: https://yashwardhan.space/work.html?p=draftncraft --- ### Wadhwani AI · 2022–2025 ## LEARN > SWAYAM · Lecture segmentation & course discovery - **Role:** Associate ML Scientist I - **Company:** Wadhwani AI · Education - **Period:** 2024 – 2025 - **Partners:** Ministry of Education India, SWAYAM, Bill & Melinda Gates Foundation Hour-long STEM lectures are hard to navigate. On **LEARN**, my primary Wadhwani focus, I helped ship lecture segmentation and course search on **SWAYAM**, turning videos into chapter clips and making courses easier to find, in a national pilot with IIT Kanpur and the Ministry of Education. ## Problem **SWAYAM** hosts thousands of hour-long STEM lectures. Without chapter markers, students scrub blindly and drop off. **LEARN** needed to split videos into topic clips and ship them on the Ministry of Education platform. **Take home:** **Lecture segmentation toolkit, deployed nationally on SWAYAM.** ![LEARN program on SWAYAM, Wadhwani AI](https://yashwardhan.space/assets/images/work/learn/learn-hero.webp) *source: Wadhwani AI* ## What I did ### AVLectures research stack - Owned the segmentation research line for LEARN: replicated [Unsupervised Audio-Visual Lecture Segmentation](https://arxiv.org/abs/2210.16644) (WACV 2023) on SWAYAM lecture corpora, clip-level **visual** (frame CNN + 3D CNN), **transcript** (BERT), and **slide OCR** cues aligned via self-supervised narration–video matching. - Fine-tuned embeddings on Indian STEM lectures, then clustered with **TW-FINCH** for temporally contiguous chapter boundaries. Lecture-domain fine-tuning on AVLectures beat zero-shot pretraining on generic video corpora. - Evaluation on educator-labeled **Courses-with-Segmentation (CwS)** subsets: **F1** and **IoU** for segment overlap, **BS@30** for boundary accuracy within 30 seconds, plus **NMI** and mean-over-frames for clustering quality. ### IIT Kanpur intern squad - Managed a parallel literature track: IIT Kanpur interns implemented **LGSS** (CVPR 2020) and **NeighborNet** on MovieNet + AVLectures, reporting mAP / mIoU against our shared benchmark suite. Owned scoping, experiment review, and synthesis. - First sponsored field visit outside Delhi **for me**, campus interviews, squad coordination, and direct alignment with IIT Kanpur stakeholders on segmentation deliverables. ### Production toolkit (ASR + LLM) - Shipped the national deployment path: **ASR transcript** → **LLM-proposed logical topic boundaries** → boundary validation against manually annotated chapters → chapter clips indexed into LEARN navigation. - Packaged as an end-to-end **lecture segmentation toolkit** for SWAYAM ingestion. Ran head-to-head evals of frontier LLMs vs. AVLectures-derived heuristics on the same held-out labeled set. - **National pilot: 10 SWAYAM courses** with chapterized lecture navigation from the toolkit output. ### Hybrid search index - Built production **hybrid retrieval** for course discovery: Elasticsearch **BM25** + dense embeddings with **reciprocal rank fusion** over titles, descriptions, and syllabi, live on SWAYAM's public discovery surface. - Led metadata ingestion pipelines and recall@k / human relevance evals as the MoE catalog expanded. ![Lecture segmentation pipeline: AVLectures research, intern literature track, ASR plus LLM production](https://yashwardhan.space/assets/images/work/learn/segmentation-pipeline.svg?v=2) *Research track (AVLectures, my implementation) and production track (ASR + LLM, nationally deployed). Intern squad ran LGSS / NeighborNet in parallel.* ## Deployment The segmentation toolkit went from AVLectures prototypes to a **10-course national pilot** on **SWAYAM**. Chapter clips fed LEARN navigation and search. The course discovery engine shipped on the same platform stack. - **10** pilot courses - **National Pilot** SWAYAM deployment - **Live** course discovery engine ## Team - [Arvind Balachandrasekaran](https://www.linkedin.com/in/arvind-balachandrasekaran-5653303) (Machine Learning Scientist · Reporting manager) - [Makarand Tapaswi](https://www.linkedin.com/in/makarand-tapaswi-9a058532) (Principal ML Scientist · Project manager) ## Outcomes - Built a **lecture segmentation toolkit** for a **10-course national pilot** on SWAYAM (AVLectures research → ASR + LLM production) - Shipped the **course discovery engine** (BM25 + embeddings + RRF) for LEARN search - Managed **IIT Kanpur intern squad** on LGSS / NeighborNet literature benchmarks under shared evaluation harness ## Sources - [Wadhwani AI, LEARN](https://www.wadhwaniai.org/impact/education-solutions/learn/) - [AVLectures (WACV 2023)](https://arxiv.org/abs/2210.16644) - [AVLectures project page](https://cvit.iiit.ac.in/research/projects/cvit-projects/avlectures) HTML case study: https://yashwardhan.space/work.html?p=learn --- ## ORF > Vachan Samiksha · Oral Reading Fluency - **Role:** Associate ML Scientist I - **Company:** Wadhwani AI · Education - **Period:** 2024 – 2025 - **Partners:** Government of Gujarat, Bill & Melinda Gates Foundation Teachers across large classes need more than a stopwatch and a single score. I contributed to the analysis layer on **Vachan Samiksha (ORF)**, turning child reading audio into **word-level errors**, reading rate, and where a student needs help, as part of a statewide program in **Gujarat government schools**. ## Problem Under **NIPUN Bharat**, teachers must assess oral reading fluency across large classes, often with a stopwatch and manual word counts. That misses *why* a child struggles: phoneme slips, substitutions, skipped text, or slow but accurate decoding. **Vachan Samiksha (ORF)** replaces the manual workflow with ASR on a student's reading audio, aligned to the passage they were given. **Take home:** Partners: **Government of Gujarat** (Sarva Shiksha Abhiyan, Vidya Samiksha Kendra) · **Bill & Melinda Gates Foundation** · deployed via **G-Shala** and **ConveGenius** in Gujarat, later **Rajasthan**. ![Oral Reading Fluency program, Wadhwani AI](https://yashwardhan.space/assets/images/work/orf/orf-hero.webp) *source: Wadhwani AI* ## What I did ### Lexical–sublexical analysis - Built alignment and scoring experiments on ASR transcripts, separating **whole-word errors** from **phoneme-level decoding mistakes** in Gujarati reading audio. - Measured how substitutions, hesitations, and statistically odd phoneme sequences show up in classroom recordings. ### Fluency metrics - Evaluated **character error rate**, **word error rate**, reading duration, and **words-correct-per-minute** as proxies for the program's fluency rubric. - Worked with the team to connect raw ASR output to teacher-facing signals, not just a single score. ### Gujarati ASR for child speech - Fine-tuned and benchmarked **Wav2Vec2-style models** on custom Gujarati child-speech data from educational reading scenarios. - Stress-tested pipelines on noisy field audio, the conditions ORF runs in across Gujarat classrooms. ## About ORF **Vachan Samiksha (ORF)** is a statewide Wadhwani AI program delivered by multiple teams over several years. The figures below are **program-wide totals** for the full deployment. Per [Wadhwani AI program data (Jan 2026)](https://www.wadhwaniai.org/programs/oral-reading-fluency/), it runs across **Grades 2–8 government schools** in **Gujarat and Rajasthan**. I contributed experiments on lexical–sublexical analysis and Gujarati child-speech ASR on field reading recordings. - **15M+** program assessments - **7.9M+** students reached - **270k+** teachers trained ## Team - Makarand Tapaswi (Principal ML Scientist · Reporting manager) - Ayush Deva (Associate ML Scientist II · ORF / Vachan Samiksha lead) - Vivek Pandey (Associate ML Scientist I · ML team) ## Outcomes - Built **lexical–sublexical ASR analysis** experiments on Gujarati child-speech reading audio - Benchmarked **Wav2Vec2-style models** on field recordings for fluency scoring (CER, WER, WCPM) ## Sources - [Wadhwani AI, Oral Reading Fluency](https://www.wadhwaniai.org/programs/oral-reading-fluency/) - [Vaachan Samiksha, program overview](https://www.wadhwaniai.org/vaachan-samiksha-leveraging-ai-to-bridge-the-literacy-divide/) HTML case study: https://yashwardhan.space/work.html?p=orf --- ## PROS > Pregnancy Risk & Outcome Stratification - **Role:** Associate ML Scientist I - **Company:** Wadhwani AI · Jhpiego - **Period:** 2023 – 2024 - **Partners:** Jhpiego Frontline nurses track hundreds of pregnancies at once. PROS reads visit history and flags who is likely to miss the next antenatal appointment, then ships the model as plain Java that runs on state health devices. ## Problem Frontline nurses (ANMs) track hundreds of pregnancies at once. A static checklist cannot say who needs a phone call **this week**. PROS reads each woman's visit history and estimates how likely she is to miss her next antenatal appointment. The score appears inside the software health workers already use, as a prioritized follow-up list. **Take home:** Shipped a **Java missed-visit risk scorer** to **Jhpiego** · promoted from **intern to Associate ML Scientist I** · first rollout in **Madhya Pradesh**. - **650+** workers trained - **6** pilot states - **MP** first rollout ![PROS risk stratification UI in antenatal workflow](https://yashwardhan.space/assets/images/work/pros/pros-hero.webp) *source: Wadhwani AI* ## Methods ### Getting the data right - Pulled antenatal visit exports from state programs, visit dates, vitals, demographics, and whether past appointments were kept. - Cleaned messy records: invalid codes, duplicate patient IDs, visits logged on impossible dates. - Built review dashboards with **Jhpiego** to catch gaps in volume and field quality before we trained anything. ### Building the model - Turned each pregnancy into a timeline, visits completed, days since the last check-up, vitals over time, prior missed visits. - Trained a **tree-based model** to predict whether the next scheduled visit would be missed. - Split data by patient cohort over time so test sets did not overlap with training pregnancies. - Chose a risk cutoff with the program team so flagged cases matched how many follow-up calls ANMs could actually make. ### Explaining scores with SHAP - Used **SHAP** to show which factors drove each risk score and which mattered most across the cohort. - Reviewed results with **Jhpiego** and clinical partners, reworked features that did not hold up on the ground. - Shared summary charts for sign-off before handing code to the software team. ![ANM using PROS during antenatal visit](https://yashwardhan.space/assets/images/work/pros/pros-anm.webp) *source: Wadhwani AI* ## Deployment State health systems could not run Python in the field. They needed the scorer as **Java on the JDK stack**, and we did not have a reliable ONNX path for our tree model at the time. ### From Python to Java - Exported the trained model with **[m2cgen](https://github.com/Bayes-Intel/m2cgen)** into readable if-else Java the integration team could audit. - Checked every test case against the Python training pipeline, scores had to match before **Jhpiego** could ship. - Documented how each input field was encoded so the software team could reproduce scores in production. ### Rollout - Worked with software engineering on alert hooks, colour-coded risk bands, and the prioritized due list. - Pilot design with **Vaibhav Sapra** and **Harsh Vats**; **Mihir Kulkarni** led the ML work. - After handoff, the module went live in **Madhya Pradesh**; other states followed. ![PROS pilot geography across Indian states](https://yashwardhan.space/assets/images/work/pros/pros-map.webp) *source: Wadhwani AI* ## Team - [Mihir Kulkarni](https://www.linkedin.com/in/mihir-kulkarni-11594aa2) (ML Scientist · team lead · Reporting manager) - [Priyanshu Mahajan](https://www.linkedin.com/in/priyanshu-mahajan) (Associate ML Scientist I · ML team) - [Vaibhav Sapra](https://www.linkedin.com/in/vaibhav-sapra-a3b95416b) (AI Product Manager · Product) - [Harsh Vats](https://www.linkedin.com/in/harsh-vats-programofficer) (Associate Program Manager · Programs · MNCH) ## Impact The scorer runs on state-provided devices during antenatal visits. Wadhwani AI reports **650+ frontline workers trained** across **six pilot states**, with **Madhya Pradesh** districts first. - Owned the full path: clean data → model → SHAP review → Java handoff to Jhpiego - Promoted from intern to **Associate ML Scientist I** when the module shipped - Confirmed field use in **Madhya Pradesh** after delivery ## Sources - [Wadhwani AI, PROS](https://www.wadhwaniai.org/impact/healthcare-solutions/pros/) - [Jhpiego](https://www.jhpiego.org/) - [m2cgen](https://github.com/Bayes-Intel/m2cgen) - [SHAP](https://github.com/shap/shap) HTML case study: https://yashwardhan.space/work.html?p=pros --- ### Projects & Hackathons · 2024 ## FedEm · Project Seshu > Decentralized foundation-model training - **Role:** Core contributor - **Company:** The Fifth Elephant · Open Source AI Hackathon - **Period:** Jan – Apr 2024 - **Hackathon:** Meta (prize sponsor), The Fifth Elephant - **Award:** National winner · ₹1,00,000 **FedEm** lets communities fine-tune language models without centralizing all the data. Adapters train locally, pass through a hub, and merge into shared checkpoints. **National winner** at **The Fifth Elephant** Open Source AI Hackathon (**₹1,00,000**, Meta-sponsored). [FedEm demo · Project Seshu at The Fifth Elephant Open Source AI Hackathon](https://yashwardhan.space/assets/videos/fedem/fedem-demo-web.mp4) ## Problem Training foundational models still concentrates GPU load in a few datacenters. Most contributors cannot pretrain at scale, yet open models like LLaMA 2 sparked demand for **community-built, multilingual alternatives**, especially for Indic languages where tokenization and compute costs bite harder. **Take home:** Category: **AI for multilingual** · Demo Day **12 Apr 2024** · **national winner** and **₹1,00,000** prize (Meta-sponsored) at [The Fifth Elephant Open Source AI Hackathon](https://hasgeek.com/hack5/open-source-ai-hackathon-2024/), one of five winning teams from 40 applications. ![The Fifth Elephant Open Source AI Hackathon 2024 winners, Project Seshu](https://yashwardhan.space/assets/images/work/fedem/fifth-elephant-hackathon.png) *The Fifth Elephant Open Source AI Hackathon 2024 · Meta-sponsored ₹1,00,000 prizes · Project Seshu national winner* ## What I did ### Continuous Relay Finetuning (CRF) - Clients train **LoRA adapters** locally, push to a HuggingFace hub, and the next client continues finetuning, relay-style instead of one centralized run. - Adapter **locking scheduler** so two clients never finetune the same adapter at once; hub checks every five minutes. - Merged adapters feed back into a foundational checkpoint with transparent version tracking. ### Seshu, multilingual SLM track - Explored **Mamba**-based pretraining for compute-efficient sequence modeling on Indic corpora (e.g. Samantar). - Server-side continuous pretraining releases checkpoints; community adapters refine against each refresh. - Client scripts for local pretrain → SFT → adapter push on the HuggingFace **mlsquare** org. ### Open source - Shipped end-to-end pipeline in the **[fedem](https://github.com/mlsquare/fedem)** PyPI package (`pip install fedem`). - Presented live demo at Hasura, Bangalore, 10 finalist teams from 40 applications. ![Continuous Relay Finetuning diagram](https://yashwardhan.space/assets/images/work/fedem/crf.png) *Continuous Relay Finetuning (CRF) · from the FedEm repo* ## Outcomes FedEm remains open source under MIT license. Initial contributors: [Soma Dhavala](https://github.com/dhavala), [Sai Nikhilesh Reddy](https://github.com/SaiNikhileshReddy), and me, community experiment to train models **by the people, for the people**. - **₹1L** national winner prize - **Top 5** of 40 teams - **PyPI** fedem package ## Team - [Soma Dhavala](https://www.linkedin.com/in/somasdhavala) (Advisor · mlsquare) - [Sai Nikhilesh Reddy](https://github.com/SaiNikhileshReddy) (Core contributor · FedEm) - [Yashwardhan Chaudhuri](https://www.linkedin.com/in/yashwardhan-chaudhuri) (Core contributor · FedEm) ## Impact Hackathon demo, federated adapter training across volunteer GPUs; checkpoints on HuggingFace/mlsquare. - National winner, The Fifth Elephant Open Source AI Hackathon 2024, ₹1,00,000 (Meta-sponsored) - FedEm library on [GitHub](https://github.com/mlsquare/fedem) and PyPI - Multilingual Mamba + LoRA roadmap for community SLM training ## Sources - [FedEm on GitHub](https://github.com/mlsquare/fedem) - [FedEm docs](https://mlsquare.github.io/fedem/) - [Full demo video](https://drive.google.com/file/d/1cknBGBbAuwkPNMwy88yZ6hylL14ErVsd/view?usp=sharing) - [Fifth Elephant hackathon](https://hasgeek.com/hack5/open-source-ai-hackathon-2024/) - [Project Seshu submission](https://hasgeek.com/hack5/open-source-ai-hackathon-2024/sub/project-seshu-SrAMqmj7TXQrLrP16kvWNT) HTML case study: https://yashwardhan.space/work.html?p=seshu --- ### Research · 2022–2024 ## Usable Security Group > IIIT Delhi · Research Associate - **Role:** Research Associate - **Company:** IIIT Delhi · Usable Security Group - **Period:** Jan 2022 – Jun 2024 With [Prof. Arun Balaji Buduru](https://faculty.iiitd.ac.in/~arunb/) in the **Usable Security Group**, I built computer vision and audio systems where outputs had to help people, not just score well on benchmarks. [Crowd scene and predicted density map](https://yashwardhan.space/assets/videos/csrnet_real_readable.mp4) ## Problem The [Usable Security Group](https://faculty.iiitd.ac.in/~arunb/) at IIIT Delhi builds **user-centric security solutions**: systems operators can actually use for public safety, health, and ecology. I joined as a **Research Associate** while still an undergrad: PyTorch pipelines, experiment design, grant writing, ablation studies, and conference papers under Prof. Arun Balaji Buduru, with **Adel Alshamrani** (University of Jeddah) co-advising the FGA crowd-counting line. **Take home:** Four publications from the lab: **IJCNN 2024**, **arXiv** lightweight fusion, and two **Interspeech 2024** audio papers, plus a side **.PY Security** CCTV prototype with usability testing. ![Crowd scene and predicted density map](https://yashwardhan.space/assets/images/work/usable-security/fga-crowd-density.png) *Density-map crowd counting, input scene and predicted heatmap (FGA paper, Fig. 1)* ## What I did ### FGA, Fourier-Guided Attention (IJCNN 2024) - Crowd density maps need **full-scale global patterns** (wide-area crowding) and **local head-level cues**. CNN-only spatial attention under-captures the global half, FGA fixes that with a **dual-path plug-in** for the last layers of CSRNet and CANNet. - **Global path:** 2D FFT on feature maps → frequency-domain processing → spatial attention. **Local path:** dilated convolutions + **channel-wise attention** on the remaining maps. Outputs are fused before the density regression head. - Benchmarks (MAE): CSRNet+FGA **8.4** on ShanghaiTech-B (baseline **10.6**); CANNet+FGA **6.9** (baseline **7.8**). Evaluated on ShanghaiTech-A/B, UCF-CC-50, and JHU++ with consistent MSE gains. - I built the **PyTorch training and data-loading pipeline**, ran architecture ablations, **Grad-CAM** qualitative analysis for interpretability, helped secure grant funding, and co-authored the paper with **Adel Alshamrani** (University of Jeddah). ![FGA module architecture diagram](https://yashwardhan.space/assets/images/work/usable-security/fga-module.png) *Dual-path FGA module, FFT global branch + conv/attention local branch (IJCNN 2024, Fig. 2)* ### Lightweight Feature Fusion (arXiv 2401.05968) - Aerial and edge deployment cannot carry VGG/ResNet-scale backbones. We kept a shared downstream counting head and swapped in **MobileNet** (**ASFNet-S**) and **MobileViT** (**ASFNet-B**) as pre-trained backbones. - **Adjacent Feature Fusion (AFF):** multi-scale feature maps from the PTM are fused across neighbouring scales before regression, same head, far fewer parameters and FLOPs than heavy SOTA. - On ShanghaiTech-A/B and UCF-CC-50, ASFNet variants hit **comparable MAE/MSE to heavier models** while staying the most compute-efficient in our comparison table. Ran **pruning** and full **ablation** sweeps for the aerial counting use case. ![ASFNet adjacent feature fusion pipeline](https://yashwardhan.space/assets/images/work/usable-security/asfnet-fusion.png) *Adjacent Feature Fusion across MobileNet / MobileViT scales (arXiv 2401.05968, Fig. 1)* ### ASGIR, Interspeech 2024 - **Problem:** ecologists need bird ID from field recordings plus habitat context for field use. - **Model:** **Audio Spectrogram Transformer (AST)** embeddings fed to an **SVM** for species classification on [Xeno-Canto](https://xeno-canto.org/) vocalizations. - **UI flow (two steps):** (1) record or upload audio → classify species; (2) enter location → **location-filtered Wikipedia scrape** for habitat and behaviour context on the predicted bird. - European subset, **51 classes**: median **100%** precision, recall, and F1 on held-out evaluation. Shipped as a usable prototype. ### VoxMed, Interspeech 2024 - **Problem:** frontline clinicians need a fast respiratory screen from stethoscope audio in one screen. - **Model:** **AST** feature extractor + **1-D CNN** classifier on the [ICBHI](https://bhichallenge.med.auth.gr/ICBHI_2017_Challenge) lung-sound corpus (recordings from Greece and Portugal). - **UI:** one-step upload → Healthy / COPD / other respiratory conditions in seconds. **90%** accuracy on the three-class task. - Ablation vs. **Wav2Vec2**, **UniSpeech**, and **WavLM**: AST won on this dataset. Built with the lab as a UI-assisted screening prototype. ### .PY Security (side project) - Side exploration with the lab: a **.PY Security** prototype for **real-time violence detection** on smart-city CCTV, motivated by limits of passive camera monitoring during mass incidents. - **SCVD** (Smart-City CCTV Violence Detection) benchmark for training; **VGG16** frame encoder + **GRU** temporal head, **BCE loss** on violence labels. Parallel **crowd-density branch** (dilated conv + spatial/channel attention, **MSE loss**) for disobedience planning. - **Operator UI:** simulated login → **Delhi map** with dummy camera pins; feeds turn **green** (peaceful) or **red** (violence flag) with **per-camera probability**. **Density view** overlays crowd estimates so planners can allocate resources during mass disobedience. - Usability sessions with security professionals and non-technical users: **density view rated most useful** for crowd-control scenarios; map interface judged intuitive, but **multi-step flows** (login → map → feed → density) needed trimming for field use. ![.PY Security violence alert dashboard with density map](https://yashwardhan.space/assets/images/work/usable-security/pysec-violence-dashboard.png) *Violence probability per CCTV feed plus density view for crowd-control planning (IIIT Delhi, 2023)* ## Outcomes Left the lab with four submitted papers, production-minded PyTorch pipelines, and grant-funded project experience before graduation. Crowd-counting work continued separately through a **Chanakya fellowship at IIT Patna** for aerial embedded deployment. - **4** papers - **IJCNN** FGA · 2024 - **2×** Interspeech · 2024 ## Team - [Arun Balaji Buduru](https://faculty.iiitd.ac.in/~arunb/) (Advisor · Founding Head · Usable Security Group) - Adel Alshamrani (Co-advisor · University of Jeddah · FGA) ## Impact Research, density-map counting with ONNX export path for aerial footage; ASGIR and VoxMed as UI-assisted audio prototypes. - [FGA: Fourier-Guided Attention Network](https://arxiv.org/abs/2407.06110), IJCNN 2024 - [Lightweight Feature Fusion](https://arxiv.org/abs/2401.05968) for resource-constrained counting - [ASGIR](https://arxiv.org/abs/2407.18927) and [VoxMed](https://arxiv.org/abs/2407.18926), Interspeech 2024 ## Sources - [USG · Prof. Arun Balaji Buduru](https://faculty.iiitd.ac.in/~arunb/) - [FGA, IJCNN 2024](https://arxiv.org/abs/2407.06110) - [Lightweight Feature Fusion](https://arxiv.org/abs/2401.05968) - [ASGIR, Interspeech 2024](https://arxiv.org/abs/2407.18927) - [VoxMed, Interspeech 2024](https://arxiv.org/abs/2407.18926) - [ASGIR code](https://github.com/MainSample1234/AS-GIR) - [VoxMed code](https://github.com/Sample-User131001/VoxMed) HTML case study: https://yashwardhan.space/work.html?p=usable-security --- ## SeekSuspect > Partial-memory suspect recognition - **Role:** Research Intern - **Company:** MIDAS Lab · IIIT Delhi - **Period:** 2022 – 2023 Delhi Police often starts from partial witness memory. Built a ranking prototype that surfaces candidates from incomplete facial descriptions for human review. Witnesses give partial features, jawline, hair, age, not mugshots. Standard face recognition assumes complete probes; operational review needs ranked candidates without false-positive escalation. ## Problem Delhi Police investigations often start from **partial witness memory**: jawline, hair texture, approximate age. Standard face matchers assume complete probes; a bad rank-one hit creates operational risk. **Take home:** Research prototype with **Delhi Police** and MIDAS Lab; ranked candidates for human review. ## What I did ### Partial-feature embeddings - Built embedding methods for incomplete facial descriptions and sketch-like witness inputs. - Ranking pipeline from partial probes to a candidate gallery with confidence thresholds for analyst review. ### Evaluation with stakeholders - Co-designed evaluation protocol with Delhi Police, what counts as a useful shortlist vs. a dangerous false lead. - Documented failure modes and review thresholds before any field consideration. ## Team - Devansh Gupta (PhD · USC Viterbi · Collaborator) - Shikhar Sharma (Collaborator) ## Impact Research prototype with law-enforcement stakeholders, not production deployment. - Partial-memory recognition methodology documented with MIDAS Lab - Operational failure modes and review thresholds defined with Delhi Police ## Sources - [MIDAS Lab · IIIT Delhi](https://midas.iiitd.edu.in/) HTML case study: https://yashwardhan.space/work.html?p=seeksuspect --- ## Experience ### Co-Founder (Product + Research) · Noteweave · evolved from ResXiv **Jan 2026 – Present** · Delhi, India - Designed E3 evaluation protocol for automated paper critique (arXiv:2605.27072): 90.2% partial-inclusive recall on 4,598 post-cutoff ICLR 2026 issue rows against GPT, Claude, and human review baselines. - Shipped LangGraph agents with FastAPI serving to 150+ users on VS Code and web app, reducing time from research brief to cited, executable plans. - Built hybrid search over 3M+ arXiv papers grounding agent and automated critique outputs. ### Founder in Residence · Founders, Inc. · Canopy Accelerator (Noteweave) **Apr 2026 – May 2026** · San Francisco, CA - Canopy residency for Noteweave: shipped LangGraph agent pipelines, hybrid search, and E3 evaluation harnesses from research prototype to production release. ### Co-Founder (Product + Tech) · ResXiv **Sep 2025 – Mar 2026** · Delhi, India - Built agentic search across 11 scholarly indexes (280M+ catalog records) with provider normalization, discipline tagging, deduplication, and PDF ingest. - Built LangGraph research agents with tool routing over search, OpenAlex, project libraries, and paper-specific RAG; streamed structured field reports through FastAPI. - Implemented cited paper chat and multi-pass draft critique over chunked PDFs with mandatory citations, OpenAlex checks, and structured issue/risk outputs. - Designed FastAPI inference paths, PostgreSQL metadata, Qdrant vector indexing, async PDF ingest, and LaTeX workspace for 300+ MAU production beta. ### Independent Consultant (Part-Time) · ML & Search Systems · DraftNCraft (Global IP Services) **Jan 2026 – Mar 2026** · Remote - Designed prior-art search architecture for 1M+ patent records at Global IP Services: hybrid lexical + semantic search over Elasticsearch and Qdrant with PostgreSQL canonical store across USPTO/EPO/WIPO filings. - Built the search gateway, indexing pipeline, and dual-index backbones for million-scale prior-art discovery. ### Associate ML Scientist I · Wadhwani AI **Oct 2023 – Sep 2025** · New Delhi, India - LEARN / SWAYAM: lecture segmentation and hybrid course search on Ministry of Education SWAYAM; developed for 2 courses in the national pilot (AVLectures research → ASR + LLM production). - ORF / Vachan Samiksha: lexical–sublexical analysis and Wav2Vec2-style Gujarati child-speech ASR on field transcripts (CER, WER, WCPM) for statewide reading fluency. - PROS (Jhpiego): tree-ensemble missed-visit risk model for antenatal care with SHAP interpretability; exported to Java via m2cgen for JDK edge deployment; field-validated with Jhpiego in Madhya Pradesh. ### Research Associate · Usable Security Group @ IIIT-Delhi **Jan 2022 – Jun 2024** · New Delhi, India - Worked under Prof. Arun Balaji Buduru on grant-funded usable-security research; developed crowd-counting architectures with 80% FLOPs reduction and 25% benchmark improvement, co-authoring FGA (IJCNN'24). - Built PyTorch training pipelines, ablations, and Grad-CAM analysis for Fourier-guided attention and lightweight feature fusion papers; contributed to grant-writing and technical proposal material. - Benchmarked five pretrained audio encoders for deepfake detection; XLS-R achieved lowest equal-error rate on ASVspoof 2019 and In-the-Wild corpora. - IJCNN 2024 technical program committee reviewer. ### CHANAKYA Research Fellow · IIT Patna **Sep 2022 – Mar 2023** · Patna, Bihar - Engineered aerial crowd-density pipeline in PyTorch/OpenCV with ONNX export on NVIDIA Jetson Nano; adapted lightweight density-map regressors for aerial crowd-counting evaluation (INR 100,000 fellowship). ## Education - **B.Tech, Computer Science with minors in Human Centered Design**, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi) ## Publications - [E3: Issue-Level Backtesting for Automated Research Critique](https://arxiv.org/abs/2605.27072), arXiv 2026. A backtesting framework for automated paper critique: 90.2% issue recall on post-cutoff ICLR submissions against GPT, Claude, and human-review baselines. - [VoxMed: Respiratory Disease from Stethoscope Audio](https://github.com/Sample-User131001/VoxMed), Interspeech 2024 - [ASGIR: Audio Spectrogram Transformer for Birds](https://github.com/MainSample1234/AS-GIR), Interspeech 2024 - [FGA: Fourier-Guided Attention Crowd Counting](https://arxiv.org/abs/2407.06110), IJCNN 2024 - [Lightweight Feature Fusion for Crowd Counting](https://arxiv.org/abs/2401.05968), arXiv 2024 --- Canonical URL: https://yashwardhan.space/ This file is the full static dump. Per-project splits: https://yashwardhan.space/llms.txt